US 11,748,595 B2
Convolution acceleration operation method and apparatus, storage medium and terminal device
Liqiang Yi, Guangdong (CN)
Assigned to Shenzhen Intellifusion Technologies Co., Ltd., Guangdong (CN)
Appl. No. 17/623,886
Filed by Shenzhen Intellifusion Technologies Co., Ltd., Guangdong (CN)
PCT Filed Oct. 27, 2020, PCT No. PCT/CN2020/124097
§ 371(c)(1), (2) Date Dec. 30, 2021,
PCT Pub. No. WO2021/088688, PCT Pub. Date May 14, 2021.
Claims priority of application No. 201911082617.1 (CN), filed on Nov. 7, 2019.
Prior Publication US 2022/0366213 A1, Nov. 17, 2022
Int. Cl. G06N 3/04 (2023.01); G06F 11/34 (2006.01)
CPC G06N 3/04 (2013.01) [G06F 11/3419 (2013.01); G06F 11/3442 (2013.01)] 7 Claims
OG exemplary drawing
 
1. A convolution acceleration operation method comprising:
obtaining convolution parameters of a target convolution layer and equipment parameters of a target hardware equipment, wherein the target convolution layer is any convolution layer of a preset convolutional neural network, and the preset convolutional neural network is to be operated in the target hardware equipment;
pre-estimating a running time of performing convolution operation on the target convolution layer in the target hardware equipment by using each of a plurality of preset acceleration algorithms, according to the convolution parameters and the equipment parameters;
determining a target acceleration algorithm corresponding to the target convolution layer and target acceleration parameters corresponding to the target acceleration algorithm, according to the running time; and
performing convolution acceleration operation on the target convolution layer by using the target acceleration algorithm and the target acceleration parameters corresponding to the target acceleration algorithm;
wherein the step of pre-estimating the running time of performing convolution operation on the target convolution layer in the target hardware equipment by using each of the plurality of preset acceleration algorithms, according to the convolution parameters and the equipment parameters, comprises:
determining a cutting mode corresponding to each of the plurality of preset acceleration algorithms, according to the convolution parameters and each of the plurality of preset acceleration algorithms;
cutting the target convolution layer by using the cutting mode corresponding to each of the plurality of preset acceleration algorithms, to obtain a plurality of convolution blocks corresponding to the cutting mode;
determining a first convolution operation amount and a first data transportation amount of performing convolution operation on each convolution block by using each of the plurality of preset acceleration algorithms, and determining a second convolution operation amount and a second data transportation amount of performing convolution operation on the target convolution layer by using each of the plurality of preset acceleration algorithms, according to the first convolution operation amount and the first data transportation amount;
determining calculation intensity of performing convolution operation on the target convolution layer by using each of the plurality of preset acceleration algorithms, according to the second convolution operation amount and the second data transportation amount; and
determining actual computation power of performing convolution operation in the target hardware equipment by using each of the plurality of preset acceleration algorithms, according to the equipment parameters and the calculation intensity, and pre-estimating the running time of performing convolution operation on the target convolution layer in the target hardware equipment by using each of the plurality of preset acceleration algorithms, according to the actual computation power and the second convolution operation amount.